RAG-NMI — Medical Device Regulations Q&A

A Retrieval-Augmented Generation (RAG) application that answers questions about medical device regulations using your own PDF documents as the knowledge base. Powered by ChromaDB for vector storage and Claude (Anthropic) as the language model.

---

How It Works

PDF documents
     │
     ▼
[ingest.py] — load → split into chunks → embed → store in ChromaDB
                                                        │
                                                        ▼
[question.py] — user query → retrieve top chunk → build prompt → Claude → answer

1. Ingestion (ingest.py) — reads all PDFs from the data/ folder, splits them into overlapping text chunks, and stores them in a local ChromaDB collection.
2. Querying (question.py) — takes a user question, finds the most relevant chunk from ChromaDB, passes it to Claude as context, and prints the answer.

Claude is instructed to answer only from the retrieved data — it will not use its own training knowledge or make things up.

---

Project Structure

rag-nmi/
├── data/               # Place your PDF documents here
├── chroma_db/          # ChromaDB persistent storage (auto-created on first ingest)
├── ingest.py           # Ingestion pipeline: PDF → chunks → ChromaDB
├── question.py         # Query pipeline: user input → RAG → Claude → answer
├── requirements.txt    # Python dependencies
├── .env                # API key (not committed)
└── README.md

---

Setup

1. Clone and create a virtual environment

git clone <repo-url>
cd retrieval-augmented-generation
python -m venv venv

# Windows
venv\Scripts\activate

# macOS / Linux
source venv/bin/activate

2. Install dependencies

pip install -r requirements.txt

Windows note: If you get a DLL load failed error for onnxruntime, install the Microsoft Visual C++ Redistributable and then run pip install --force-reinstall onnxruntime.

3. Set your Anthropic API key

Create a .env file in the project root:

ANTHROPIC_API_KEY=your-api-key-here

Get your key at console.anthropic.com.

4. Add your documents

Place PDF files containing medical device regulation content in the data/ folder.

---

Usage

Step 1 — Ingest documents

Run this once (or whenever you add new PDFs):

python ingest.py

This loads all PDFs, splits them into 300-character chunks with 100-character overlap, and stores them in chroma_db/.

Step 2 — Ask questions

python question.py

You will be prompted to enter a question. The app retrieves the most relevant chunk and asks Claude to answer based only on that context.

---

Example

What do you want to know about Medical device regulations?

What are MDR and IVDR?


---------------------

Based on the information provided, I can tell you that:

- **MDR** refers to the Medical Devices Regulation
- **IVDR** refers to the In Vitro Diagnostic (Medical Devices) Regulation

However, the data provided only mentions these terms in the context of guidance
documents being prepared by the European Commission and codes that match for
manufacturers of medical devices. It does not provide a detailed definition or
explanation of what these regulations specifically entail.

So beyond knowing what the abbreviations stand for, I don't have enough
information to give you more details about these regulations.

The answer is deliberately limited to what appears in the retrieved chunk. If the source PDFs contain more detail about MDR/IVDR, a broader retrieval (n_results > 1) would surface it.

---

Configuration

Variable	File	Description
DATA_PATH	ingest.py / question.py	Folder containing source PDFs
CHROMA_PATH	ingest.py / question.py	ChromaDB storage directory
chunk_size	ingest.py	Characters per chunk (default: 300)
chunk_overlap	ingest.py	Overlap between chunks (default: 100)
n_results	question.py	Number of chunks retrieved per query (default: 1)
model	question.py	Claude model used (default: claude-opus-4-8)
max_tokens	question.py	Maximum tokens in Claude's response (default: 1024)

---

Dependencies

Package	Purpose
anthropic	Claude API client
chromadb	Local vector database for storing and retrieving chunks
langchain_community	PyPDFDirectoryLoader for reading PDFs
langchain_text_splitters	RecursiveCharacterTextSplitter for chunking
pypdf	PDF parsing backend used by LangChain
python-dotenv	Loads ANTHROPIC_API_KEY from .env
onnxruntime	Required by ChromaDB's default embedding model

---

Limitations

• Only the top 1 chunk is retrieved by default. For questions that span multiple sections of a document, increase n_results in question.py and update the prompt accordingly.
• Chunk size (300 characters) is small. Larger chunks provide more context per retrieval but may dilute relevance. Tune chunk_size and chunk_overlap in ingest.py to suit your documents.
• The ChromaDB collection name is medicinsk_regelverk. If you switch document sets, either drop and recreate the collection or change the name in both files.
• Re-running ingest.py uses upsert, so existing IDs are overwritten rather than duplicated.